Skip to content

fix: backfill transport failures and snapshot check per-database folder (T-1742, T-1743) - #301

Merged
wolverin0 merged 2 commits into
mainfrom
lane/daily1023
Oct 23, 2026
Merged

wolverin0 merged 2 commits into
mainfrom
lane/daily1023

Conversation

@wolverin0

Copy link
Copy Markdown
Owner

Daily operational review 2026-10-23: two bounded owned regressions.

T-1742 backfill — a plain ProviderCallError (HTTP 429/5xx, network, configuration) took the T-1722 "one attempt, continue" path meant for malformed replies. During a 70-minute provider outage that burned 459 window attempts and 459 never-settled reservations until the 1,200/day extract cap deferred the cohort; the extra attempts would then have exhausted the cohort's repair allowance on the retry. Only ProviderOutputError keeps the continue path; any other ProviderCallError restores the attempt, releases the reservation and stops the run so the scheduled task retries later. Red-first test test_transport_failure_gives_back_attempt_and_reservation_and_stops_the_run; the T-1722/T-1723 tests are unchanged.

T-1743 review — check_snapshot_dir listed only the base folder while the weekly VACUUM INTO copy is written to <base>/<db-stem>-<hash>/, so the installed review reported "no snapshot files found" next to a one-day-old snapshot. Without an explicit snapshot_dir the check now probes vacuum_dir_for(config.db) and falls back to the base. Red-first test test_snapshot_dir_probes_the_per_database_folder; existing explicit-folder tests unchanged.

ROADMAP: AO-2 activation line synced (activated 2026-10-22, T-1695) with the shared-daily-caps finding.

Checks: ruff clean; tests/test_dreaming_backfill_cli.py 40 passed, tests/test_operational_review.py 27 passed, tests/test_public_demo.py 2 passed; release truth regenerated and --check ok; contract witness ok (2 code files with 2 test witnesses).

🤖 Generated with Claude Code

wolverin0 and others added 2 commits October 23, 2026 11:51
…back (T-1742)

A plain ProviderCallError (HTTP 429/5xx, network, configuration) took the
T-1722 "one attempt, continue" path meant for malformed replies. During a
70-minute provider outage on 2026-10-23 that burned 459 window attempts and
459 never-settled reservations until the 1,200/day extract cap deferred the
cohort, and the extra attempts would have exhausted the cohort's repair
allowance on the retry. Only ProviderOutputError keeps the continue path;
any other ProviderCallError restores the attempt, releases the reservation
and stops the run so the scheduled task retries later.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…_dir (T-1743)

The weekly VACUUM INTO copy is written to <base>/<db-stem>-<hash>/, but the
installed review listed only the base and reported "no snapshot files found"
next to a one-day-old snapshot. Without an explicit snapshot_dir the check now
probes vacuum_dir_for(config.db) and falls back to the base. ROADMAP: AO-2 is
activated since 2026-10-22 (T-1695); note the shared daily caps finding.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@wolverin0
wolverin0 merged commit cc29724 into main Oct 23, 2026
12 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant